def blend_images(x, progress, resolution_schedule, num_blocks):
"""Blends images of different resolutions according to `progress`.
When training `progress` is at a stable stage for resolution r, returns
image `x` downscaled to resolution r and then upscaled to `final_resolutions`,
call it x'(r).
Otherwise when training `progress` is at a transition stage from resolution
r to 2r, returns a linear combination of x'(r) and x'(2r).
Args:
x: An image `Tensor` of NHWC format with resolution `final_resolutions`.
progress: A scalar float `Tensor` of training progress.
resolution_schedule: An object of `ResolutionSchedule`.
num_blocks: An integer of number of blocks.
Returns:
An image `Tensor` which is a blend of images of different resolutions.
"""
x_blend = []
for block_id in range(1, num_blocks + 1):
alpha = _generator_alpha(block_id, progress)
scale = resolution_schedule.scale_factor(block_id)
x_blend.append(alpha *
layers.upscale(layers.downscale(x, scale), scale))
return tf.add_n(x_blend)
def blend_images(x, progress, resolution_schedule, num_blocks):
"""Blends images of different resolutions according to `progress`.
When training `progress` is at a stable stage for resolution r, returns
image `x` downscaled to resolution r and then upscaled to `final_resolutions`,
call it x'(r).
Otherwise when training `progress` is at a transition stage from resolution
r to 2r, returns a linear combination of x'(r) and x'(2r).
Args:
x: An image `Tensor` of NHWC format with resolution `final_resolutions`.
progress: A scalar float `Tensor` of training progress.
resolution_schedule: An object of `ResolutionSchedule`.
num_blocks: An integer of number of blocks.
Returns:
An image `Tensor` which is a blend of images of different resolutions.
"""
x_blend = []
for block_id in range(1, num_blocks + 1):
alpha = _generator_alpha(block_id, progress)
scale = resolution_schedule.scale_factor(block_id)
x_blend.append(alpha * layers.upscale(layers.downscale(x, scale), scale))
return tf.add_n(x_blend)
def test_upscale_4d_images_returns_upscaled_images(self):
x_np = np.array([[[[1, 2, 3]]], [[[4, 5, 6]]]], dtype=np.float32)
with self.test_session(use_gpu=True) as sess:
x1_np, x2_np = sess.run(
[layers.upscale(tf.constant(x_np), n) for n in [1, 2]])
expected2_np = [[[[1, 2, 3], [1, 2, 3]], [[1, 2, 3], [1, 2, 3]]],
[[[4, 5, 6], [4, 5, 6]], [[4, 5, 6], [4, 5, 6]]]]
self.assertNDArrayNear(x1_np, x_np, 1.0e-5)
self.assertNDArrayNear(x2_np, expected2_np, 1.0e-5)
def test_blend_images_in_transition_stage(self):
x_np = np.random.normal(size=[2, 8, 8, 3])
x = tf.constant(x_np, tf.float32)
x_blend = networks.blend_images(
x,
tf.constant(0.2),
resolution_schedule=networks.ResolutionSchedule(scale_base=2,
num_resolutions=2),
num_blocks=2)
with self.test_session(use_gpu=True) as sess:
x_blend_np = sess.run(x_blend)
x_blend_expected_np = 0.8 * sess.run(
layers.upscale(layers.downscale(x, 2), 2)) + 0.2 * x_np
self.assertNDArrayNear(x_blend_np, x_blend_expected_np, 1.0e-6)
def generator(z,
progress,
num_filters_fn,
resolution_schedule,
num_blocks=None,
kernel_size=3,
colors=3,
to_rgb_activation=None,
scope='progressive_gan_generator',
reuse=None):
"""Generator network for the progressive GAN model.
Args:
z: A `Tensor` of latent vector. The first dimension must be batch size.
progress: A scalar float `Tensor` of training progress.
num_filters_fn: A function that maps `block_id` to # of filters for the
block.
resolution_schedule: An object of `ResolutionSchedule`.
num_blocks: An integer of number of blocks. None means maximum number of
blocks, i.e. `resolution.schedule.num_resolutions`. Defaults to None.
kernel_size: An integer of convolution kernel size.
colors: Number of output color channels. Defaults to 3.
to_rgb_activation: Activation function applied when output rgb.
scope: A string or variable scope.
reuse: Whether to reuse `scope`. Defaults to None which means to inherit
the reuse option of the parent scope.
Returns:
A `Tensor` of model output and a dictionary of model end points.
"""
if num_blocks is None:
num_blocks = resolution_schedule.num_resolutions
start_h, start_w = resolution_schedule.start_resolutions
final_h, final_w = resolution_schedule.final_resolutions
def _conv2d(scope, x, kernel_size, filters, padding='SAME'):
return layers.custom_conv2d(
x=x,
filters=filters,
kernel_size=kernel_size,
padding=padding,
activation=lambda x: layers.pixel_norm(tf.nn.leaky_relu(x)),
he_initializer_slope=0.0,
scope=scope)
def _to_rgb(x):
return layers.custom_conv2d(x=x,
filters=colors,
kernel_size=1,
padding='SAME',
activation=to_rgb_activation,
scope='to_rgb')
end_points = {}
with tf.variable_scope(scope, reuse=reuse):
with tf.name_scope('input'):
x = tf.contrib.layers.flatten(z)
end_points['latent_vector'] = x
with tf.variable_scope(block_name(1)):
x = tf.expand_dims(tf.expand_dims(x, 1), 1)
x = layers.pixel_norm(x)
# Pad the 1 x 1 image to 2 * (start_h - 1) x 2 * (start_w - 1)
# with zeros for the next conv.
x = tf.pad(
x, [[0] * 2, [start_h - 1] * 2, [start_w - 1] * 2, [0] * 2])
# The output is start_h x start_w x num_filters_fn(1).
x = _conv2d('conv0', x, (start_h, start_w), num_filters_fn(1),
'VALID')
x = _conv2d('conv1', x, kernel_size, num_filters_fn(1))
lods = [x]
for block_id in range(2, num_blocks + 1):
with tf.variable_scope(block_name(block_id)):
x = layers.upscale(x, resolution_schedule.scale_base)
x = _conv2d('conv0', x, kernel_size, num_filters_fn(block_id))
x = _conv2d('conv1', x, kernel_size, num_filters_fn(block_id))
lods.append(x)
outputs = []
for block_id in range(1, num_blocks + 1):
with tf.variable_scope(block_name(block_id)):
lod = _to_rgb(lods[block_id - 1])
scale = resolution_schedule.scale_factor(block_id)
lod = layers.upscale(lod, scale)
end_points['upscaled_rgb_{}'.format(block_id)] = lod
# alpha_i is used to replace lod_select. Note sum(alpha_i) is
# garanteed to be 1.
alpha = _generator_alpha(block_id, progress)
end_points['alpha_{}'.format(block_id)] = alpha
outputs.append(lod * alpha)
predictions = tf.add_n(outputs)
batch_size = z.shape[0].value
predictions.set_shape([batch_size, final_h, final_w, colors])
end_points['predictions'] = predictions
return predictions, end_points
def test_upscale_invalid_scale_throws_exception(self):
with self.assertRaises(ValueError):
self.assertRaises(layers.upscale(tf.constant([]), -1))
def generator(z,
progress,
num_filters_fn,
resolution_schedule,
num_blocks=None,
kernel_size=3,
colors=3,
to_rgb_activation=None,
scope='progressive_gan_generator',
reuse=None):
"""Generator network for the progressive GAN model.
Args:
z: A `Tensor` of latent vector. The first dimension must be batch size.
progress: A scalar float `Tensor` of training progress.
num_filters_fn: A function that maps `block_id` to # of filters for the
block.
resolution_schedule: An object of `ResolutionSchedule`.
num_blocks: An integer of number of blocks. None means maximum number of
blocks, i.e. `resolution.schedule.num_resolutions`. Defaults to None.
kernel_size: An integer of convolution kernel size.
colors: Number of output color channels. Defaults to 3.
to_rgb_activation: Activation function applied when output rgb.
scope: A string or variable scope.
reuse: Whether to reuse `scope`. Defaults to None which means to inherit
the reuse option of the parent scope.
Returns:
A `Tensor` of model output and a dictionary of model end points.
"""
if num_blocks is None:
num_blocks = resolution_schedule.num_resolutions
start_h, start_w = resolution_schedule.start_resolutions
final_h, final_w = resolution_schedule.final_resolutions
def _conv2d(scope, x, kernel_size, filters, padding='SAME'):
return layers.custom_conv2d(
x=x,
filters=filters,
kernel_size=kernel_size,
padding=padding,
activation=lambda x: layers.pixel_norm(tf.nn.leaky_relu(x)),
he_initializer_slope=0.0,
scope=scope)
def _to_rgb(x):
return layers.custom_conv2d(
x=x,
filters=colors,
kernel_size=1,
padding='SAME',
activation=to_rgb_activation,
scope='to_rgb')
end_points = {}
with tf.variable_scope(scope, reuse=reuse):
with tf.name_scope('input'):
x = tf.contrib.layers.flatten(z)
end_points['latent_vector'] = x
with tf.variable_scope(block_name(1)):
x = tf.expand_dims(tf.expand_dims(x, 1), 1)
x = layers.pixel_norm(x)
# Pad the 1 x 1 image to 2 * (start_h - 1) x 2 * (start_w - 1)
# with zeros for the next conv.
x = tf.pad(x, [[0] * 2, [start_h - 1] * 2, [start_w - 1] * 2, [0] * 2])
# The output is start_h x start_w x num_filters_fn(1).
x = _conv2d('conv0', x, (start_h, start_w), num_filters_fn(1), 'VALID')
x = _conv2d('conv1', x, kernel_size, num_filters_fn(1))
lods = [x]
for block_id in range(2, num_blocks + 1):
with tf.variable_scope(block_name(block_id)):
x = layers.upscale(x, resolution_schedule.scale_base)
x = _conv2d('conv0', x, kernel_size, num_filters_fn(block_id))
x = _conv2d('conv1', x, kernel_size, num_filters_fn(block_id))
lods.append(x)
outputs = []
for block_id in range(1, num_blocks + 1):
with tf.variable_scope(block_name(block_id)):
lod = _to_rgb(lods[block_id - 1])
scale = resolution_schedule.scale_factor(block_id)
lod = layers.upscale(lod, scale)
end_points['upscaled_rgb_{}'.format(block_id)] = lod
# alpha_i is used to replace lod_select. Note sum(alpha_i) is
# garanteed to be 1.
alpha = _generator_alpha(block_id, progress)
end_points['alpha_{}'.format(block_id)] = alpha
outputs.append(lod * alpha)
predictions = tf.add_n(outputs)
batch_size = z.shape[0].value
predictions.set_shape([batch_size, final_h, final_w, colors])
end_points['predictions'] = predictions
return predictions, end_points